Quantitative analysis method and system for gastric cancer claudin 18

Abstract

The present application relates to the technical field of gastric cancer, and discloses a tumor cell quantification analysis method and system based on gastric cancer CLDN18, which comprises the following steps: presetting candidate points on a pathological section image; generating foreground and background auxiliary points for cell labeling points, and constructing a structured supervised target; extracting multi-scale semantic features through a pre-trained pathological large model, and generating an aggregated feature map through downsampling alignment and fusion operations; adaptively reconstructing candidate point features using implicit feature interpolation technology; predicting position offset and class confidence based on candidate point features, and determining cell position and class through threshold screening and offset optimization; optimizing the model in combination with an auxiliary point supervision mechanism, and outputting objective quantification results. The present application can significantly improve the positioning accuracy and classification consistency of CLDN18 positive tumor cells, and reduce the workload of doctors.

Description

Technical Field

[0001] This invention belongs to the field of gastric cancer technology, and particularly relates to a method and system for quantitative analysis of tumor cells based on gastric cancer CLDN18. Background Technology

[0002] In recent years, CLDN18 targeted therapy has made significant progress in the treatment of gastric cancer patients. Specifically, CLDN18 protein expression has been proven to be a key biomarker for screening potential beneficiaries, and its expression is closely related to tumor occurrence, development, and prognosis. In clinical practice for gastric cancer, CLDN18 immunohistochemical expression detection has become an important means of assessing the applicability of anti-CLDN18 targeted therapy. The so-called CLDN18 immunohistochemical expression detection assessment refers to the process of searching for and counting 2+ / 3+ positive tumor cells and 0 / 1+ negative tumor cells on a whole gastric cancer CLDN18 immunohistochemical pathological section, and accurately calculating the proportion of positive tumor cells to all tumor cells.

[0003] However, with the rapid development of CLDN18.2 targeted therapy, its pathological testing, as the "gold standard" for patient screening, faces a severe challenge in terms of consistency. Currently, the performance of CLDN18.2 immunohistochemical (IHC) testing varies significantly in the real world, directly affecting the accurate formulation of treatment plans and the clinical benefits to patients. At present, the evaluation and analysis of CLDN18 immunohistochemical pathological slides mainly relies on manual image analysis. Professional pathologists need to search for suspicious tissue areas one by one under a microscope, and then complete the final score estimation based on their accumulated professional knowledge. However, there are significant differences in the test results between different laboratories, especially the false negative (FN) rate (12%) is significantly higher than the false positive (FPR, 1%), especially in borderline positive cases. This situation may lead to a large number of patients who could benefit from targeted therapy being wrongly excluded, which deserves high clinical attention. Currently, cell detection methods based on deep learning are mainly divided into two categories: (1) cell detection methods based on bounding box annotation; (2) cell detection methods based on point annotation.

[0004] 1) Box-Annotation-Based Cell Detection Quantization Methods: Box-annotation-based cell detection methods often use object detection networks for cell quantification and statistics. Object detection networks typically encode multi-scale features and their contextual information using convolutional neural networks. The encoded deep features are then used with region generation networks to obtain candidate region boxes. Finally, techniques such as non-maximum suppression and fully connected classification are used to regress the category and location of the candidate region boxes, thereby achieving cell detection and classification. However, box-annotation-based cell detection methods also have shortcomings. On the one hand, box annotation is costly; on the other hand, although they perform well in sparse abnormal cell detection scenarios, their performance is often poor in dense cell detection scenarios.

[0005] 2) Cell detection methods based on point annotation: In pathological images with dense cell distribution, point annotation saves a lot of costs and effectively alleviates the annotation workload of clinicians. Cell detection methods based on point annotation can be further subdivided into two types according to the implementation process: (1) Cell detection methods based on semantic segmentation networks; (2) Cell detection methods based on point regression. Among them, the cell detection method based on semantic segmentation networks predicts the cell distribution density map of the pathological image at the pixel level through image segmentation technology, converts the density map into a binary map, selects the center point as the cell detection position, and then performs cell quantification statistics. The cell detection method based on point regression performs quantification statistics on cells in the pathological image through an end-to-end target detection network. Specifically, based on point annotation information, the regression branch and classification branch in the target detection network are used to predict the offset of the location points of the dense population and identify the target category, respectively, and the Hungarian algorithm is used to match the distance between the prediction result and the real annotation, thereby effectively locating cells. However, unlike other modal medical images, the cell features in pathological images are messy and confusing, the feature differences of different types of cells are relatively small, and there is a lack of separable boundaries. During training, the instability of matching target candidate points with real labeled points often leads to unclear learning objectives for the network, resulting in missed detections and false detections.

[0006] Therefore, there is an urgent need to develop methods and systems for quantitative analysis of tumor cells based on CLDN18 in gastric cancer. These methods could significantly improve the accuracy of CLDN18-positive tumor cell localization and classification consistency, reducing the workload of physicians. The consistency challenge in CLDN18.2 pathological detection remains a major obstacle to achieving precision medicine. Therefore, the application of artificial intelligence-assisted interpretation will gradually improve the consistency and reliability of CLDN18.2 detection. Only by ensuring the accuracy and reproducibility of the detection can the potential of CLDN18.2 targeted therapy be fully realized, allowing more gastric cancer patients to benefit from precision medicine. Summary of the Invention

[0007] To address the aforementioned technical issues, this invention provides a method and system for quantitative analysis of tumor cells based on CLDN18 in gastric cancer, which can significantly improve the localization accuracy and classification consistency of CLDN18-positive tumor cells and reduce the workload of doctors.

[0008] This invention provides a method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer, the method comprising the following steps:

[0009] S1. Pre-determine several candidate points covering the entire image on the CLDN18 pathological slide image of gastric cancer immunohistochemistry.

[0010] S2. Label each cell in the immunohistochemical CLDN18 pathological slide image of gastric cancer, and define the auxiliary points corresponding to each cell labeling point. Construct a target point set for each cell labeling point and its corresponding auxiliary points. The labeling content includes the coordinates and category of each cell.

[0011] S3. Multi-scale semantic features of the annotated pathological slide images are extracted and aggregated using a pre-trained large-scale pathological model to obtain an aggregated feature map.

[0012] S4. Based on the aggregated feature map, perform implicit feature interpolation on each candidate point to obtain the reconstructed features of the candidate points;

[0013] S5. Based on the reconstruction features of each candidate point, predict the position offset of the candidate point relative to the target point and the category confidence.

[0014] S6. Filter candidate points with a category confidence level greater than or equal to a preset threshold, and determine the location and type of each cell based on the position offset and category confidence level of the filtered candidate points.

[0015] S7. Output the quantitative results of tumor cells based on the location and type of each cell.

[0016] Furthermore, in S2, the auxiliary points include foreground auxiliary points and background auxiliary points:

[0017] Foreground auxiliary points are randomly offset within the first preset interval of the cell annotation points. Generate within;

[0018] Background auxiliary points are randomly offset within a second preset range of cell annotation points. ]∪[ Generate within;

[0019] Where, n neg Much greater than n pos .

[0020] Furthermore, in S3, multi-scale semantic features of the labeled pathological slide images are extracted and aggregated using a pre-trained large-scale pathology model to obtain an aggregated feature map, including:

[0021] S31. Use a pre-trained large pathological model to extract four-layer multi-scale semantic features from the labeled pathological slide images.

[0022] S32. Downsample the features of the first three layers and align the spatial resolution of the features of the first three layers with the spatial resolution of the features of the fourth layer.

[0023] S33. By fusing the four layers of features through feature addition operations, an aggregated feature map is obtained.

[0024] Furthermore, in S4, based on the aggregated feature map, implicit feature interpolation is performed on each candidate point to obtain the reconstructed features of the candidate points, including:

[0025] S41. Determine four different auxiliary dimensions, and obtain four integer multiples based on the four different auxiliary dimensions and the size of the aggregated feature map; where the auxiliary dimension is the size that is closest to the size of the original gastric cancer immunohistochemical CLDN18 pathological slide image and can be divided evenly by the size of the aggregated feature map.

[0026] S42. Divide the coordinates of each candidate point by four integer multiples to obtain the four neighboring feature points of the candidate point on the aggregated feature map.

[0027] S43. Use Fourier transform to encode the position of each neighboring feature point to obtain position encoding information;

[0028] S44. Calculate the reconstructed features of the candidate points based on the feature vectors, positional encoding information, and relative distances between each neighboring feature point and its corresponding candidate point.

[0029] Furthermore, in S44, based on the feature vectors, positional encoding information, and relative distances between each neighboring feature point and its corresponding candidate point, the reconstructed features of the candidate point are calculated, as shown in the following expression:

[0030] ;

[0031] in, This represents the reconstructed features of candidate points. f mlp Z represents the parameters of the multilayer perceptron, p represents the p-th neighboring feature point, and Z represents the multilayer perceptron parameters. p Let Q represent the eigenvector of the p-th neighboring feature point. p V represents the relative distance between the p-th neighboring feature point and its corresponding candidate point. p This represents the location encoding information of the p-th neighboring feature point.

[0032] Furthermore, in S2, the auxiliary points corresponding to each cell annotation point also include:

[0033] Calculate the matching loss for foreground auxiliary points and background auxiliary points separately;

[0034] The matching loss of the foreground auxiliary point is calculated based on the spatial distance between the foreground auxiliary point and the candidate point and the classification confidence of the foreground auxiliary point. It is used to enable the model to learn the regression and confidence classification of candidate points in the neighborhood of the labeled point.

[0035] The matching loss calculation process for foreground auxiliary points is as follows: For each cell in the image, iterate through all its foreground auxiliary points and all candidate points, accumulate the sum of the logarithm of the classification confidence of each foreground auxiliary point and the weighted first spatial distance deviation, and perform normalization based on the total number of cells, the number of foreground auxiliary points, and the number of candidate points; wherein, the first spatial distance deviation is the sum of the absolute differences between the horizontal and vertical coordinates between the foreground auxiliary point and the candidate point, and multiplied by the first loss weight hyperparameter;

[0036] The matching loss of background auxiliary points is calculated based on the spatial distance between background auxiliary points and candidate points and the classification confidence of background auxiliary points. It is used to suppress interfering candidate points and their confidence that are far away from the labeled points.

[0037] The process of calculating the matching loss of background auxiliary points is as follows: For each cell in the image, iterate through all its background auxiliary points and all candidate points, accumulate the logarithm of the classification confidence complement of each background auxiliary point and the sum of the weighted second spatial distance deviation, and perform normalization based on the total number of cells, the number of background auxiliary points and the number of candidate points; where the second spatial distance deviation is the sum of the absolute differences between the horizontal and vertical coordinates between the background auxiliary point and the candidate point, and multiplied by the second loss weight hyperparameter.

[0038] Furthermore, the formula for calculating the matching loss of the foreground auxiliary points is as follows:

[0039] ;

[0040] in, The foreground auxiliary point matching loss is represented by N, which represents the number of cells in the CLDN18 pathological section image of gastric cancer immunohistochemistry. K represents the number of foreground auxiliary points. p Let l represent the number of candidate points, l represent the l-th cell, and i represent the i-th foreground auxiliary point. This represents the classification confidence of the i-th foreground auxiliary point of the l-th cell. This represents the first loss weight hyperparameter. This represents the x-axis coordinate of the i-th foreground auxiliary point of the l-th cell. This represents the y-axis coordinate of the i-th foreground auxiliary point of the l-th cell. This represents the x-axis coordinate of the k-th candidate point in the l-th cell. This represents the y-coordinate of the k-th candidate point in the l-th cell;

[0041] The formula for calculating the matching loss of background auxiliary points is as follows:

[0042] ;

[0043] in, This represents the matching loss of background auxiliary points. This represents the number of background auxiliary points, where j represents the j-th background auxiliary point. This represents the classification confidence of the j-th background auxiliary point in the l-th cell. This represents the second loss weight hyperparameter. This represents the x-axis coordinate of the j-th background auxiliary point of the l-th cell. This represents the y-axis coordinate of the j-th background auxiliary point of the l-th cell.

[0044] Furthermore, following S6, the model is optimized by fusing the real annotation loss with the matching loss of foreground and background auxiliary points. The calculation formula is as follows:

[0045] ;

[0046] ;

[0047] ;

[0048] in, Represents the overall loss function. Represents the regression loss function. Represents the classification loss function. T represents the third loss weight hyperparameter. l P represents the true coordinates of the l-th cell. l C represents the predicted coordinates of the l-th cell. l Y represents the prediction confidence level for the l-th cell. l This indicates the category label for the l-th cell.

[0049] This invention also provides a tumor cell quantification analysis system based on gastric cancer CLDN18, used to perform the above-mentioned tumor cell quantification analysis method based on gastric cancer CLDN18. The system includes the following modules:

[0050] The candidate point annotation module is used to pre-define several candidate points covering the entire image on the CLDN18 pathological slide image of gastric cancer immunohistochemistry.

[0051] The auxiliary point annotation module is used to annotate each cell in the CLDN18 pathological slide image of gastric cancer immunohistochemistry, and to define the auxiliary points corresponding to each cell annotation point. The annotation content includes the coordinates and category of each cell.

[0052] The feature encoding module is used to extract multi-scale semantic features from labeled pathological slide images through a pre-trained large-scale pathological model and aggregate them to obtain an aggregated feature map.

[0053] The implicit feature interpolation module is used to perform implicit feature interpolation on each candidate point based on the aggregated feature map to obtain the reconstructed features of the candidate points;

[0054] The candidate point prediction module is used to predict the position offset of each candidate point relative to the target point and the category confidence based on the reconstruction features of each candidate point.

[0055] The cell prediction module is used to filter candidate points with a category confidence level greater than or equal to a preset threshold, and to determine the position and type of each cell based on the position offset.

[0056] The output module is used to output the quantification results of tumor cells based on the location and type of each cell.

[0057] The embodiments of the present invention have the following technical effects:

[0058] This invention achieves global coverage detection of pathological slides by pre-setting uniformly distributed candidate points, replacing the traditional manual region-by-region observation mode, significantly shortening the evaluation time and improving the consistency of results. It utilizes a foreground and background auxiliary point generation mechanism to guide the model to focus on the effective cell region and ignore interfering background, improving the accuracy of identifying densely overlapping areas and reducing missed and false detections. It employs implicit feature interpolation technology to adaptively reconstruct sub-pixel-level features, overcoming positional deviations caused by image scaling and improving the reliability of small-sized cell localization. Combined with multi-scale feature aggregation and joint loss optimization models, it improves prediction accuracy, providing objective evidence for clinical targeted therapy while reducing the workload of physicians. Attached Figure Description

[0059] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0060] Figure 1 This is a flowchart of a tumor cell quantification analysis method based on gastric cancer CLDN18 provided in an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of the framework of the refined intelligent assessment method for gastric cancer immunohistochemistry CLDN18 provided in the embodiments of the present invention;

[0062] Figure 3 This is a schematic diagram of the implicit feature interpolation principle provided in the embodiments of the present invention;

[0063] Figure 4 This is a comparison chart of qualitative results of the tumor cell quantitative analysis method based on CLDN18 of gastric cancer provided in the embodiments of the present invention; Figure 4 (a) is an immunohistochemical pathological image of CLDN18 gastric cancer; Figure 4 (b) shows the clinical annotation results; Figure 4 (c) shows the prediction results of this method. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0065] Claudin18.2 (CLDN18.2), a member of the tight junction protein family, is expressed at low levels in normal gastric mucosa, but its epitope is exposed after malignant transformation, resulting in high expression in tumors such as gastric cancer and pancreatic cancer, with an expression rate of approximately 38%, and exhibiting heterogeneity (reaching 48.3% in diffuse gastric cancer). Targeted drugs such as zotuximab have shown significant efficacy in phase III trials and have been recommended by guidelines such as NCCN and CSCO. Therefore, CLDN18.2 immunohistochemical detection has become a key biomarker for screening patients who will benefit, but its evaluation still relies on manual analysis, facing challenges in consistency. Large-scale EQA studies show that the false negative rate among laboratories reaches 12%, especially borderline positive cases, which are prone to misinterpretation, leading to missed diagnoses. Artificial intelligence-assisted interpretation can improve objectivity through deep learning, but it relies on precise annotation. Current methods, such as semantic segmentation or point regression networks, still suffer from missed and false positives due to feature confusion and unstable matching. Therefore, there is an urgent need to develop high-precision quantitative analysis methods.

[0066] To address the aforementioned issues, this invention provides a method and system for quantitative analysis of tumor cells based on CLDN18 in gastric cancer. By using foreground / background auxiliary points and implicit feature interpolation, the method significantly improves the localization accuracy and classification consistency of CLDN18 positive cells, reduces the manual workload, and provides an objective basis for targeted therapy.

[0067] This invention provides a method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer. Figure 1 This is a flowchart of a tumor cell quantification analysis method based on CLDN18 in gastric cancer provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of the refined intelligent assessment method for gastric cancer immunohistochemistry CLDN18 provided in this embodiment of the invention. See also... Figure 1 and Figure 2 The method includes the following steps:

[0068] S1. Pre-determine several candidate points covering the entire image on the CLDN18 pathological slide image of gastric cancer immunohistochemistry.

[0069] Current CLDN18.2 immunohistochemical (IHC) testing faces consistency challenges. Large-scale studies show that inter-laboratory differences lead to false negative rates as high as 12%, particularly affecting borderline positive cases and potentially causing patients to miss targeted therapy. Traditional manual interpretation is highly subjective, while AI-assisted systems can improve objectivity through deep learning, but rely on precise annotation. This invention overcomes the limitations of manual region-by-region searching by pre-setting global candidate points (step S1), laying the foundation for subsequent automated analysis.

[0070] In some embodiments, the CLDN18 pathological slide image of gastric cancer immunohistochemistry is first divided into several image blocks, and a predetermined number of fixed grid points (e.g., 2000) are uniformly set on the image blocks as candidate points to cover the potential cell areas of the entire image as basic detection units, thus avoiding manual searching of each region.

[0071] S2. Label each cell in the immunohistochemical CLDN18 pathological slide image of gastric cancer, and define the auxiliary point corresponding to each cell label point. Construct a target point set by combining each cell label point and its corresponding auxiliary point.

[0072] The annotations include the coordinates and category of each cell.

[0073] In some embodiments, the auxiliary points include foreground auxiliary points. and background auxiliary points :

[0074] Foreground auxiliary points are randomly offset within the first preset interval of the cell annotation points. Generated within, for example, It can be set to 5, depending on the cell size and image scaling factor;

[0075] Background auxiliary points are randomly offset within a second preset range of cell annotation points. ]∪[ Generated within, for example, It can be set to 50;

[0076] The number of foreground and background auxiliary points is selected according to actual needs.

[0077] Furthermore, defining the auxiliary points corresponding to each cell annotation point also includes:

[0078] Calculate the matching loss for foreground auxiliary points and background auxiliary points separately;

[0079] For each generated foreground auxiliary point, the predicted offset and classification confidence are obtained based on its spatial location. The purpose of this scheme is to ensure that the classification confidence of the foreground auxiliary point is as close as possible to the true class and has a value of 1. At the same time, the predicted offset of the point is highly consistent with the added random number. The formula for calculating the matching loss of the foreground auxiliary point is as follows:

[0080] ;

[0081] in, The foreground auxiliary point matching loss is represented by N, where N represents the number of cells in the CLDN18 pathological section image of gastric cancer immunohistochemistry. K represents the number of foreground auxiliary points. p Let l represent the number of candidate points, l represent the l-th cell, and i represent the i-th foreground auxiliary point. This represents the classification confidence of the i-th foreground auxiliary point of the l-th cell. This represents the first loss weight hyperparameter. This represents the x-axis coordinate of the i-th foreground auxiliary point of the l-th cell. This represents the y-axis coordinate of the i-th foreground auxiliary point of the l-th cell. This represents the x-axis coordinate of the k-th candidate point in the l-th cell. This represents the y-coordinate of the k-th candidate point in the l-th cell;

[0082] For each generated background auxiliary point, the predicted offset and classification confidence are obtained based on its spatial location. The goal of this scheme is to ensure that the classification confidence of the background auxiliary point is as close to 0 as possible, and that the predicted offset of the point is also close to 0. The formula for calculating the matching loss of the background auxiliary point is as follows:

[0083] ;

[0084] in, This represents the matching loss of background auxiliary points. This represents the number of background auxiliary points, where j represents the j-th background auxiliary point. This represents the classification confidence of the j-th background auxiliary point in the l-th cell. This represents the second loss weight hyperparameter. This represents the x-axis coordinate of the j-th background auxiliary point of the l-th cell. This represents the y-axis coordinate of the j-th background auxiliary point of the l-th cell.

[0085] Specifically, the matching loss between foreground and background auxiliary points aims to address the unstable matching between candidate and labeled points in dense cell detection. Its design principle involves constructing a structured learning objective: foreground auxiliary points force the model to learn accurate regression and high-confidence classification of candidate points within the neighborhood of the labeled point, while background auxiliary points suppress interfering candidate points and their confidence levels that are far from the labeled point. This dual constraint mechanism provides the model with a clear direction for optimization, significantly improving the stability and accuracy of localization and classification in complex cell distribution scenarios.

[0086] Furthermore, to prevent overfitting of the deep neural network and ensure the robustness of the model, the image patch dataset is augmented online using methods such as random rotation and random shearing, and random noise is added to enhance the images online. In this embodiment, the momentum optimizer is used to optimize the classification network, with a batch size of 24, an initial learning rate of 0.007, a maximum stride of 30000, a weight decrease of 0.0002, and a momentum of 0.9, and three-fold cross-validation is performed. Simultaneously, 20% of the training set samples are reserved for monitoring the network's performance. and Set them to 4 and 16 respectively.

[0087] S3. Multi-scale semantic features of the labeled pathological slide images are extracted and aggregated using a pre-trained large-scale pathological model to obtain an aggregated feature map.

[0088] Artificial intelligence-assisted interpretation is key to resolving consistency issues in CLDN18.2 detection. AI systems can learn from expert-annotated samples to form stable standards, reducing human error. In step S3 of this invention, a pre-trained large-scale pathological model (such as UNI) is used for multi-scale feature aggregation to simulate this process, improving the robustness of feature representation and providing reliable input for cell-level analysis.

[0089] In some embodiments, S3 includes the following sub-steps:

[0090] S31. Use a pre-trained large pathology model to extract four-layer multi-scale semantic features from the labeled pathology slide images.

[0091] In this embodiment, the pre-trained large pathological model adopts the UNI model. Specifically, for the q-th image x... q The parameters of the large pathological model UNI are defined as follows: The image is then feature-encoded to capture its four-layer multi-scale semantic features Fx. q e e = [1, ..., 4].

[0092] S32. Downsample the features of the first three layers and align the spatial resolution of the features of the first three layers with the spatial resolution of the features of the fourth layer.

[0093] Since the four layers of semantic features have different spatial resolutions after encoding, feature pooling can be used to downsample the first three layers of features to make their spatial resolution consistent with that of the fourth layer.

[0094] S33. The four layers of features are fused through feature addition to obtain the aggregated feature map Fx. q .

[0095] S4. Based on the aggregated feature map, perform implicit feature interpolation on each candidate point to obtain the reconstructed features of the candidate points.

[0096] In cell detection tasks, traditional feature extraction methods rely on regular grid structures. After feature pooling, the spatial location becomes blurred, making it difficult to flexibly handle the position of candidate points, resulting in inaccurate feature representation and affecting model performance. Therefore, this invention designs an implicit feature interpolation method that models features at arbitrary locations using a multilayer perceptron, achieving adaptive feature extraction and thus improving model performance.

[0097] In some embodiments, Figure 3 This is a schematic diagram of the implicit feature interpolation principle provided in an embodiment of the present invention. See also: Figure 3 S4 includes the following sub-steps:

[0098] S41. Determine four different auxiliary dimensions, and obtain four integer multiples based on the four different auxiliary dimensions and the size of the aggregated feature map.

[0099] The auxiliary size is the size that is closest to the size of the original gastric cancer immunohistochemical CLDN18 pathological slide image and can be divided evenly by the aggregated feature map.

[0100] S42. Divide the coordinates of each candidate point by four integer multiples to obtain the four neighboring feature points of the candidate point on the aggregated feature map.

[0101] S43. Use Fourier transform to encode the location of each neighboring feature point to obtain the location encoding information.

[0102] S44. Calculate the reconstructed features of the candidate points based on the feature vectors, positional encoding information, and relative distances between each neighboring feature point and its corresponding candidate point.

[0103] In some embodiments, the expression for calculating the reconstructed features of candidate points is as follows:

[0104] ;

[0105] in This represents the reconstructed features of candidate points. f mlp Z represents the parameters of the multilayer perceptron, p represents the p-th neighboring feature point, and Z represents the multilayer perceptron parameters. p Let Q represent the eigenvector of the p-th neighboring feature point. p V represents the relative distance between the p-th neighboring feature point and its corresponding candidate point. p This represents the location encoding information of the p-th neighboring feature point.

[0106] S5. Based on the reconstruction features of each candidate point, predict the position offset of the candidate point relative to the target point and the category confidence.

[0107] In some embodiments, based on the reconstructed features of each candidate point, a P2PNet network model is used to predict the positional offset and class confidence. Specifically, the positional regression branch and the class prediction branch predict the positional offset and class confidence of the candidate points, respectively. The positional regression branch employs a coordinate offset-based regression strategy, using the candidate point feature values ​​and performing feature dimensionality reduction through residual blocks and 1×1 convolutions to ultimately generate the offset of each candidate point. The classification prediction branch uses a confidence scoring mechanism to distinguish between valid and invalid candidate points. For the feature values ​​of candidate points, feature dimensionality reduction is performed by residual blocks and 1×1 convolutions, and finally, the class confidence of each candidate point is generated.

[0108] S6. Filter candidate points with a category confidence level greater than or equal to a preset threshold, and determine the location and type of each cell based on the position offset of the filtered candidate points and the category confidence level.

[0109] In some embodiments, the preset threshold is set to 0.5 by default, but can be adjusted according to actual conditions. Candidate points with a class confidence score less than the preset threshold are discarded as negative samples, while candidate points with a class confidence score greater than or equal to the preset threshold are retained as positive samples. Among the positive samples, the coordinates and class of the candidate point with the smallest position offset and the highest confidence score are selected as the position and category of its corresponding target point.

[0110] Furthermore, after determining the location and type of each cell in S6, the model is optimized by fusing the real annotation loss with the matching loss of foreground and background auxiliary points. The calculation formula is as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] in, Represents the overall loss function. Represents the regression loss function. Represents the classification loss function. T represents the third loss weight hyperparameter. l P represents the true coordinates of the l-th cell. l C represents the predicted coordinates of the l-th cell (the predicted coordinates of a cell are the sum of the coordinates of the corresponding candidate point and the predicted position offset), C l Y represents the prediction confidence of the l-th cell (the prediction confidence of a cell is the prediction class confidence of the corresponding candidate point). l This represents the category label of the l-th cell. Both the regression loss function and the classification loss function are designed based on traditional object detection methods.

[0115] In some embodiments, we compared our tumor cell quantification analysis method based on CLDN18 in gastric cancer with a traditional approach without auxiliary points (Baseline). The quantitative comparison results are shown in Table 1:

[0116] Table 1. Quantitative Comparison Analysis Results Between This Scheme and the Traditional Scheme

[0117]

[0118] The F1 score, which considers both precision and recall, measures the overall performance of the model. The data shows that the F1 score of this method is significantly higher than the baseline, indicating higher accuracy and comprehensiveness in identifying and classifying tumor cells in CLDN18 pathological slide images of gastric cancer via immunohistochemistry. Mean squared error (MSE) is a common indicator of the difference between predicted and true values. A smaller MSE indicates a smaller difference between the predicted and true values. The MSE of this method is significantly lower than the baseline, indicating that the method has smaller errors in predicting cell location and type, and the prediction results are closer to the true values. Root mean squared error (RMSE), the square root of the mean squared error, also measures the difference between predicted and true values. The RMSE of this method is also significantly lower than the baseline, further validating its superiority in reducing prediction errors. Mean absolute percentage error (MAPE) measures the relative difference between predicted and true values. The MAPE of this method is much lower than the baseline, indicating that the method has smaller relative errors in predicting cell location and type, and the prediction results are more reliable.

[0119] Through the comparative analysis of the above quantitative results, we can conclude that:

[0120] This method outperforms the baseline in terms of F1 score, mean squared error, root mean square error, and mean absolute percentage error, demonstrating its significant advantages in the quantitative analysis of tumor cells in CLDN18 pathological slide images from gastric cancer immunohistochemistry. By introducing foreground and background auxiliary points and combining them with a pre-trained large-scale pathological model to extract multi-scale semantic features, this method effectively improves the accuracy and reliability of cell location and species prediction. Compared to the traditional baseline approach, this method can more accurately identify and classify tumor cells in CLDN18 pathological slide images from gastric cancer immunohistochemistry, providing a more reliable basis for clinical diagnosis and treatment.

[0121] CLDN18.2, as an emerging target for gastric cancer treatment, has made encouraging progress in areas such as ADC drugs, CAR-T cell therapy, and monoclonal antibodies, offering new treatment hope for patients with advanced gastric cancer. However, the consistency challenge of CLDN18.2 pathological detection remains a major obstacle to achieving precision medicine. To address the issue of highly subjective interpretation criteria, AI-assisted pathological image analysis systems may provide more objective and consistent interpretation results. These systems can use deep learning to develop stable interpretation criteria from a large number of expert-annotated samples, reducing human error. The accuracy of the annotation determines the effectiveness of the AI-assisted interpretation system. Therefore, the application of AI-assisted interpretation will gradually improve the consistency and reliability of CLDN18.2 detection. Only by ensuring the accuracy and reproducibility of the detection can the potential of CLDN18.2 targeted therapy be fully realized, allowing more gastric cancer patients to benefit from precision medicine.

[0122] Figure 4 This is a comparison chart of qualitative results of the tumor cell quantitative analysis method based on CLDN18 in gastric cancer provided in this embodiment of the invention. Figure 4 (a) is the immunohistochemical pathological image of CLDN18 in gastric cancer. Figure 4 (b) shows the clinical annotation results. Figure 4 (c) shows the prediction results of this method. Different colored dots in the figure represent different types of cells (e.g., strongly positive tumor cells, moderately positive tumor cells, weakly positive tumor cells, intestinal metaplastic cells, etc.). From the qualitative results, it can be found that the prediction results of the method proposed in this invention are highly consistent with the clinical annotations.

[0123] S7. Output the quantitative results of tumor cells based on the location and type of each cell.

[0124] Specifically, the system automatically locates CLDN18-positive tumor cells based on their location and type, calculates their proportion scores (e.g., percentage of 1+ / 2+ / 3+ positive cells), and outputs the final overall tumor cell category (positive / negative) based on the percentage of positive cells and a preset proportion threshold. Artificial intelligence-assisted medical pathology image recognition is currently mainly applied to scenarios such as statistical analysis of the proportion of positive areas and cell-level analysis. Taking cell-level analysis as an example, one typical application is PD-L1 expression interpretation, calculated as: (number of positive tumor cells + number of positive tumor-associated lymphocytes) / total number of tumor cells. In contrast, the CLDN (Claudin protein) expression assessment, which this study focuses on, is also based on the cell level, but uses a different calculation method: first, the proportion of positive tumor cells with staining intensities of 1+, 2+, and 3+ is calculated separately; then, the sum of the proportions of 2+ and 3+ positive cells is calculated as the evaluation indicator.

[0125] This invention achieves global coverage detection of pathological slides by pre-setting uniformly distributed candidate points, replacing the traditional manual region-by-region observation mode, significantly shortening the evaluation time and improving the consistency of results. It utilizes a foreground and background auxiliary point generation mechanism to guide the model to focus on the effective cell region and ignore interfering background, improving the accuracy of identifying densely overlapping areas and reducing false negatives and false positives. It employs implicit feature interpolation technology to adaptively reconstruct sub-pixel-level features, overcoming positional deviations caused by image scaling and improving the reliability of small-sized cell localization. Combined with multi-scale feature aggregation and joint loss optimization models, it achieves global detection, reduces false negatives, and improves prediction accuracy. With the development of CLDN18.2 targeted therapy, it can further improve the treatment landscape for advanced gastric cancer, providing patients with more effective treatment options for clinical targeted therapy while reducing the workload of physicians.

[0126] In summary, the development of CLDN18.2 targeted therapy relies on high-precision detection. This invention, through automated quantitative analysis, outputs the proportion of positive cells (e.g., the percentage of 2+ / 3+ cells), providing objective evidence for clinical practice. AI-assisted interpretation improves consistency and reliability; the qualitative results of this invention's embodiments are highly consistent with clinical annotations, validating this advantage and reducing the burden on physicians while promoting precision medicine.

[0127] This invention also provides a tumor cell quantification analysis system based on gastric cancer CLDN18, used to perform the above-described tumor cell quantification analysis method based on gastric cancer CLDN18. The system includes the following modules:

[0128] The candidate point annotation module is used to pre-define several candidate points covering the entire image on the CLDN18 pathological slide image of gastric cancer immunohistochemistry.

[0129] The auxiliary point annotation module is used to annotate each cell in the CLDN18 pathological slide image of gastric cancer immunohistochemistry, and to define the auxiliary points corresponding to each cell annotation point. The annotation content includes the coordinates and category of each cell.

[0130] The feature encoding module is used to extract multi-scale semantic features from labeled pathological slide images through a pre-trained large-scale pathological model and aggregate them to obtain an aggregated feature map.

[0131] The implicit feature interpolation module is used to perform implicit feature interpolation on each candidate point based on the aggregated feature map to obtain the reconstructed features of the candidate points;

[0132] The candidate point prediction module is used to predict the position offset of each candidate point relative to the target point and the category confidence based on the reconstruction features of each candidate point.

[0133] The cell prediction module is used to filter candidate points with a category confidence level greater than or equal to a preset threshold, and to determine the position and type of each cell based on the position offset.

[0134] The output module is used to output the quantification results of tumor cells based on the location and type of each cell.

Claims

1. A method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer, characterized in that, The method includes the following steps: S1. Pre-determine several candidate points covering the entire image on the CLDN18 pathological slide image of gastric cancer immunohistochemistry. S2. Label each cell in the CLDN18 pathological slide image of gastric cancer immunohistochemistry and define the auxiliary points corresponding to each cell labeling point. Construct a target point set for each cell labeling point and its corresponding auxiliary points. The labeling content includes the coordinates and category of each cell. S3. The multi-scale semantic features of the labeled pathological slide images are extracted and aggregated using a pre-trained pathological large model to obtain an aggregated feature map. S4. Based on the aggregated feature map, perform implicit feature interpolation on each candidate point to obtain the reconstructed features of the candidate points; specifically including: S41. Determine four different auxiliary dimensions, and obtain four integer multiples based on the four different auxiliary dimensions and the size of the aggregated feature map; wherein, the auxiliary dimension is the size that is closest to the size of the original gastric cancer immunohistochemical CLDN18 pathological slide image and can be divided evenly by the size of the aggregated feature map; S42. Divide the coordinates of each candidate point by four integer multiples to obtain the four neighboring feature points of the candidate point on the aggregated feature map; S43. Use Fourier transform to encode the position of each neighboring feature point to obtain position encoding information; S44. Calculate the reconstructed features of the candidate points based on the feature vectors, position encoding information, and relative distances between each neighboring feature point and its corresponding candidate point. S5. Based on the reconstruction features of each candidate point, predict the position offset and category confidence of the candidate point relative to the target point; S6. Filter candidate points whose category confidence is greater than or equal to a preset threshold, and determine the position and type of each cell based on the position offset of the filtered candidate points and the category confidence. S7. Output the quantitative results of tumor cells based on the location and type of each cell.

2. The method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer according to claim 1, characterized in that, In step S2, the auxiliary points include foreground auxiliary points and background auxiliary points: The foreground auxiliary points are randomly offset within a first preset range of the cell annotation points. Generate within; The background auxiliary points are randomly offset within a second preset range of the cell annotation points. ]∪[ Generate within; Where, n neg Much greater than n pos .

3. The method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer according to claim 1, characterized in that, In step S3, multi-scale semantic features of the labeled pathological slide images are extracted and aggregated using a pre-trained large-scale pathological model to obtain an aggregated feature map, including: S31. Use a pre-trained large pathological model to extract four layers of multi-scale semantic features from the labeled pathological slide images; S32. Downsample the features of the first three layers and align the spatial resolution of the features of the first three layers with the spatial resolution of the features of the fourth layer. S33. By fusing the four layers of features through feature addition operations, an aggregated feature map is obtained.

4. The method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer according to claim 1, characterized in that, In step S44, the reconstructed features of the candidate points are calculated based on the feature vectors, positional encoding information, and relative distances between each neighboring feature point and its corresponding candidate point. The expression is as follows: ； in, This represents the reconstructed features of candidate points. f mlp Z represents the parameters of the multilayer perceptron, p represents the p-th neighboring feature point, and Z represents the multilayer perceptron parameters. p Let Q represent the eigenvector of the p-th neighboring feature point. p V represents the relative distance between the p-th neighboring feature point and its corresponding candidate point. p This represents the location encoding information of the p-th neighboring feature point.

5. The method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer according to claim 2, characterized in that, In S2, defining the auxiliary points corresponding to each cell annotation point also includes: Calculate the matching loss for the foreground auxiliary point and the background auxiliary point respectively; The matching loss of the foreground auxiliary point is calculated based on the spatial distance between the foreground auxiliary point and the candidate point and the classification confidence of the foreground auxiliary point. It is used to enable the model to learn the regression and confidence classification of candidate points in the neighborhood of the labeled point. The matching loss calculation process for the foreground auxiliary points is as follows: For each cell in the image, iterate through all its foreground auxiliary points and all candidate points, accumulate the sum of the logarithm of the classification confidence of each foreground auxiliary point and the weighted first spatial distance deviation, and perform normalization processing based on the total number of cells, the number of foreground auxiliary points and the number of candidate points; wherein, the first spatial distance deviation is the sum of the absolute differences between the horizontal and vertical coordinates between the foreground auxiliary point and the candidate point, and multiplied by the first loss weight hyperparameter; The matching loss of the background auxiliary points is calculated based on the spatial distance between the background auxiliary points and the candidate points and the classification confidence of the background auxiliary points, and is used to suppress interfering candidate points and their confidence values ​​that are far away from the labeled points. The matching loss calculation process for the background auxiliary points is as follows: For each cell in the image, iterate through all its background auxiliary points and all candidate points, accumulate the logarithm of the classification confidence complement of each background auxiliary point and the sum of the weighted second spatial distance deviation, and perform normalization processing based on the total number of cells, the number of background auxiliary points and the number of candidate points; wherein, the second spatial distance deviation is the sum of the absolute differences between the horizontal and vertical coordinates between the background auxiliary point and the candidate point, and multiplied by the second loss weight hyperparameter.

6. The method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer according to claim 5, characterized in that, The formula for calculating the matching loss of the foreground auxiliary points is as follows: ； in, The foreground auxiliary point matching loss is represented by N, where N represents the number of cells in the CLDN18 pathological section image of gastric cancer immunohistochemistry. K represents the number of foreground auxiliary points. p Indicates the number of candidate points. Indicates the first There are _ cells, where i represents the i-th foreground auxiliary point. Indicates the first The classification confidence of the i-th foreground auxiliary point of each cell. This represents the first loss weight hyperparameter. Indicates the first The x-axis coordinate of the i-th foreground auxiliary point of each cell Indicates the first The y-axis coordinate of the i-th foreground auxiliary point of a cell Indicates the first The x-axis coordinate of the kth candidate point of each cell Indicates the first The y-axis coordinate of the kth candidate point of each cell; The formula for calculating the matching loss of the background auxiliary points is as follows: ； in, This represents the matching loss of background auxiliary points. This represents the number of background auxiliary points, where j represents the j-th background auxiliary point. Indicates the first The classification confidence of the j-th background auxiliary point of each cell. This represents the second loss weight hyperparameter. Indicates the first The x-axis coordinate of the j-th background auxiliary point of a cell. Indicates the first The y-axis coordinate of the j-th background auxiliary point of each cell.

7. The method for quantitative analysis of tumor cells based on CLDN18 in gastric cancer according to claim 6, characterized in that, Following S6, the model is further optimized by fusing the real annotation loss based on the matching loss of the foreground auxiliary points and the background auxiliary points. The calculation formula is as follows: ； ； ； in, Represents the overall loss function. Represents the regression loss function. Represents the classification loss function. T represents the third loss weight hyperparameter. l Indicates the first The true coordinates of each cell, P l Indicates the first Predicted coordinates of each cell, C l Indicates the first Prediction confidence of each cell, Y l Indicates the first Individual cell category labeling.

8. A tumor cell quantification analysis system based on gastric cancer CLDN18, used to perform the tumor cell quantification analysis method based on gastric cancer CLDN18 as described in any one of claims 1-7, characterized in that, The system includes the following modules: The candidate point annotation module is used to pre-define several candidate points covering the entire image on the CLDN18 pathological slide image of gastric cancer immunohistochemistry. The auxiliary point annotation module is used to annotate each cell in the CLDN18 pathological slide image of gastric cancer immunohistochemistry, and to define the auxiliary points corresponding to each cell annotation point. The annotation content includes the coordinates and category of each cell. The feature encoding module is used to extract multi-scale semantic features of the labeled pathological slide images through a pre-trained large pathological model and aggregate them to obtain an aggregated feature map. An implicit feature interpolation module is used to perform implicit feature interpolation on each candidate point based on the aggregated feature map to obtain the reconstructed features of the candidate points; The candidate point prediction module is used to predict the position offset and category confidence of each candidate point relative to the target point based on the reconstruction features of each candidate point. The cell prediction module is used to filter candidate points whose category confidence is greater than or equal to a preset threshold, and to determine the position and type of each cell based on the position offset. The output module is used to output the quantification results of tumor cells based on the location and type of each cell.

Images